Everyone blames the dashboard. The graph was confusing, the panel was missing, nobody could tell if the spike was normal. But most dashboards don't fail at the dashboard. They fail upstream, at the metric, long before anyone opens Grafana.

I spend most of my time on the instrumentation side, and that's the view this talk takes: a dashboard can only be as good as the metrics feeding it. A counter that should have been a histogram. A label you wish you'd added, now unqueryable without a code change. Cardinality that quietly made the whole thing slow. By the time you're arranging panels, those decisions are already made for you.

You'll leave able to look at a metric, before it's ever graphed, and tell whether it can support the dashboard you actually need.